Connor Leahy — TAIRA

Connor Leahy is an AI researcher and activist who first became prominent as a co-founder of EleutherAI, an open research collective that built and released large language models, and later founded the AI-safety company Conjecture. That path—from helping democratize access to large models to campaigning for strict limits on the most powerful systems—is central to his public argument. ControlAI, where he is listed as U.S. executive director, says his current work focuses on policy responses to the threat posed by superintelligent AI.

Leahy argues that the deepest problem is not simply malicious use or biased output, but the creation of increasingly capable software that its developers cannot reliably understand or control. In a 2024 Conjecture roadmap, he and co-author Gabe Alfour wrote that the current development path could produce systems able to automate and replace human activity while undermining cybersecurity and human agency. Their proposed alternative is “cognitive software”: systems designed to remain legible, bounded, and under meaningful human control rather than autonomous general-purpose agents.

His regulatory position is correspondingly stronger than voluntary safety commitments. In a TIME interview about deepfakes and catastrophic AI risk, Leahy argued that policy must address the supply chain—including developers and distributors—rather than punish only people who misuse finished systems. He treats regulation of present harms as both worthwhile in itself and a way to build institutions capable of governing more dangerous systems later. TIME has also reported his support for an international agreement prohibiting development above a defined capability or scale threshold because current safety evaluations cannot establish that frontier systems are controllable.

In a long-form discussion on the state of alignment research, Leahy compared repeatedly deploying more capable systems without dependable safety guarantees to playing Russian roulette. The recorded interview and transcript also captures an important distinction in his position: he does not seek a halt to all useful AI research, but argues that a moratorium on frontier capability scaling can be separated from work intended to understand, evaluate, or secure AI systems. His public case is therefore aimed at general-purpose systems capable of accumulating strategic power, not narrowly scoped tools such as scientific prediction software.

For readers concerned about AI risk, Leahy represents an unusually direct call to political action. He rejects the idea that market competition will naturally reward caution, because each developer can externalize catastrophic risk while capturing the benefits of moving faster. His preferred response is public authority with the power to monitor large training efforts, require evidence of safety, and prevent systems from being built when their creators cannot demonstrate control. Whether or not one accepts his probability estimates, his challenge is concrete: society should not treat an inability to prove safety as permission to conduct an irreversible experiment.

Sources